Speech to text
Choose a speech to text provider for your voice agent by language and latency, and tune transcriber settings such as confidence, interruptions, silence and boosted keywords.
Speech to text (STT), also called the transcriber, turns what the caller says into text for the LLM. If it mishears, the agent answers the wrong question, so pick the provider that is most accurate for your callers’ languages and accents.
Set it in the voice agent under Transcriber Settings.
Which speech to text providers are supported?
| Provider | Models | Languages |
|---|---|---|
| Deepgram | Flux General (English and multilingual), Nova 3, Nova 3 Medical, Nova 2, Nova, Enhanced and Base, several with regional variants | 54 options, including Multi, Hindi, and English (US, UK, Australia, New Zealand, India) |
| Chirp 3, Chirp 2, Latest Long, Latest SHORT | About 130 language and region options | |
| Sarvam | Sarvam Saarika | 11 Indian languages: English, Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu |
| Soniox | STT Realtime | 60 languages, plus Additional Languages for calls that mix languages |
| ElevenLabs | Scribe Realtime | About 99 languages |
Choose the Provider, then the Language and Model. With Soniox you can also pick Additional Languages as hints for secondary languages callers may switch to.
Which provider should I choose?
| Your callers speak | Recommended | Why |
|---|---|---|
| English only, lowest latency | Deepgram Nova 2 or Nova 3 | High accuracy and very low latency (under 300 ms) for English. Nova 2 is strongest on pure English. |
| English and Hindi, or a mix | Soniox, with Hindi and English set | Best accuracy for Hindi and Hinglish. Handles language switching, accents and background noise well. Nova 3 is the alternative if you prefer Deepgram. |
| Several Indian languages (for example Hindi plus Telugu, Tamil or Kannada) | Soniox or Google | Wider language coverage. Test accuracy for each specific language before choosing. |
| Hindi only, lowest latency | Sarvam | Built for Indian languages with very low latency, but more sensitive to background noise. |
Google is also a good fallback for languages the specialised providers do not cover, though it may be less accurate than them in noisy calls.
What do the transcriber settings do?
These settings change how the agent hears and when it decides the caller has finished. The recommended values are a starting point: adjust them for your callers and test the result.
| Setting | What it does | Range | Default | Recommended |
|---|---|---|---|---|
| Confidence Threshold | Transcripts below this confidence are dropped and not sent to the LLM. Raise it for noisy calls, but too high a value drops real words. | 0 to 1 | 0.4 | 0.4 to 0.5 |
| Background Denoising Enabled | Filters background noise while the caller is talking. | On or off | On | On |
| Convert Numerals | Writes spoken numbers as digits, so “five thousand” becomes “5000”. | On or off | On | On |
| Interruption min words | How many words the caller must say before the agent stops talking to listen. Shorter replies still reach the LLM, they just do not interrupt. | 1 to 10 | 2 | 2 to 3 |
| User voice seconds | How long the caller must keep speaking before it counts as an interruption. | 0.1 to 10 seconds | 0.3 | 2.5 to 3.5 seconds |
| Backoff seconds | Cool-down after the caller interrupts, before the agent starts speaking again. | 0.1 to 10 seconds in the builder. On calls, 0.2 to 0.4 seconds takes effect: lower values count as 0.2 and higher values as 0.4. | 0.5 (applied as 0.4) | 0.2 to 0.4 seconds |
| Silence seconds | How long a pause marks the end of the caller’s turn. Too short and the agent talks over people who pause mid-sentence; too long and replies feel slow. | 0.01 to 10 seconds | 0.01 | 2 to 4 seconds |
| Smart Turn Detection | Uses an AI model to judge whether the caller has finished their turn, instead of relying on silence alone. | On or off | Off | Off to start; try it if the agent often interrupts |
| Smart Turn Threshold | How sure the model must be that the turn has ended. Shown when Smart Turn Detection is on. | 0.1 to 0.95 | 0.5 | 0.5 |
| Smart Turn Max Wait | The longest the agent waits for the model before replying anyway. | 0.5 to 10 seconds | 3 | 3 |
| Boosted Keywords | Words and phrases to recognise reliably. Add your brand, product names and uncommon terms, separated by commas. | Free text | Empty | Your brand and product terms |
Can I add call-specific keywords?
Yes. When you trigger a call through the API, pass extra terms in key_terms, or set include_variable_key_terms to true to add the words in that call’s variables (such as the customer’s name) as keywords. They are added to Boosted Keywords for that call only. See Triggering calls.